Inside the NVIDIA Agent Toolkit: The Open-Source Blueprint for Autonomous Enterprise Agents

OpenShell is the secure runtime environment that sits between the agent and the underlying infrastructure. It enforces network, policy, and privacy boundaries through process-level isolation and a privacy router, ensuring agents operate within pre-defined guardrails . Under an Apache 2.0 license, OpenShell can run locally, on cloud VMs, or at the edge. Partners including Microsoft, Red Hat, and Canonical are integrating it across Windows, Linux, and cloud environments .

Nemotron 3 Ultra is the latest open-source model in the Nemotron family, purpose-built for agentic workloads. A 550-billion-parameter mixture-of-experts model, it has been post-trained to handle long-running, multi-step tasks. NVIDIA reports 5x faster inference speeds and up to 30% lower total cost of ownership compared to previous generations on complex agentic benchmarks .

CUDA-X Libraries are exposed as agent-callable skills, giving agents direct access to NVIDIA's acceleration stack. The toolkit specifically surfaces cuDF for data processing, cuOpt for mathematical optimization, PhysicsNeMo for scientific simulation, and CUDA-Q for quantum computing tasks. AI-Q blueprints additionally provide a research-agent scaffold .

Who Is Using It: Enterprise Software Adopters

NVIDIA's go-to-market strategy centers on enterprise independent software vendors (ISVs) that embed the toolkit into their platforms rather than selling directly to end users . At GTC Taipei, NVIDIA announced a roster of partners adopting the stack across three critical domains :

Chip Design and Engineering

Cadence is using OpenShell to secure its ChipStack AI Super Agent, a fully autonomous AI engineer for chip design and verification. NVIDIA confirmed it is the first customer to use ChipStack autonomously to verify its own chip designs . Synopsys and Siemens are similarly building autonomous agents for electronic design automation and industrial simulation workflows, compressing weeks of complex engineering work into hours .

Cybersecurity

CrowdStrike and Palantir are building always-on AI agents powered by Nemotron open models for threat detection, response, and operational decision-making. CrowdStrike integrates the toolkit to enable autonomous cybersecurity workflows that defend critical infrastructure across cloud, data center, and edge environments .

Enterprise Platforms and Operations

Adobe, Salesforce, SAP, and ServiceNow are among the first to integrate the toolkit into their platforms for content creation, CRM, enterprise resource planning, and IT service management . A broader list of launch partners from NVIDIA's GTC 2026 announcements includes Atlassian, Box, Cohesity, IQVIA, Amdocs, Cisco, and Red Hat .

Physical AI Agent Skills: Robots, Cars, and Digital Twins

In a parallel announcement, NVIDIA released a significant open-source collection of physical AI agent skills and tools spanning four platforms :

• NVIDIA Omniverse libraries for industrial digital twin simulation
• NVIDIA Cosmos world foundation models for physical reasoning and video generation
• NVIDIA Alpamayo tools for autonomous driving development
• NVIDIA Metropolis skills for vision AI and intelligent video analytics

These skills turn complex robotics, autonomous vehicle, and industrial simulation pipelines into agent-executable instructions—covering data generation, simulation training, evaluation, and deployment in repeatable, automated workflows . Industry leaders using these tools include Agile Robots, Delta Electronics, Foxconn, Pegatron, PTC, and TSMC, alongside many of the same names adopting the enterprise Agent Toolkit .

How the Pieces Fit Together

The strategic significance of the Agent Toolkit lies in its vertical integration. Instead of offering disconnected pieces, the toolkit provides a complete runtime for building, testing, deploying, and securing autonomous agents from start to finish—a move that positions NVIDIA's software stack as infrastructure for the agent era .